OdinDTA: Combining Mutual Attention and Pre-training for Drug-target Affinity Prediction

Accurate and effective Drug Target binding Affinity (DTA) prediction can significantly shorten the drug development lifecycle and reduce the cost. Although many deep learning-based methods have been developed for DTA prediction, most do not model complex drug-target interaction process and have poor interpretability. In addition, these models depend on large-scale labelled data. To address these problems, we designed a new DTA prediction model called OdinDTA. We use drug sequences and graphs to extract drug features in this model. To meet the challenge of labelled data scarcity, our studies adopted self-supervised pre-training tasks to learn information of amino acid sequences of proteins. Finally, we utilize the mutual attention mechanism to fuse the representations of drugs and proteins. We evaluate the performance of our method on two benchmark datasets, KIBA and Davis. Experimental results show that our model outperforms the current state-of-the-art methods on two independent datasets.